This paper implements a machine learning(ML)-based procedure for constructing the missing sensor(s) data in a net zero energy building in case of complete failure in data recording (for up to one hour). In the first scenario, missing temperature data is re-created using the sensor's ex-ante data, the HVAC system's status flag, and the ambient conditions. In the second scenario, the temperature data (until failure occurred) from two close-by spaces are also utilized as inputs. For each scenario, ML-based pipelines' performance is first assessed by considering different prediction horizons using a benchmark algorithm. Next, each pipeline's most promising features and the most suitable algorithm are identified. Using the obtained optimal pipeline, a sliding window-based training scheme is implemented, and the size of the training window is optimized. It is shown that feature selection, algorithm optimization procedures, and the sliding window-based training scheme notably improve the forecasting performance. The proposed methodology can be deployed as a tool in intervals with total data logging failure, providing data to ML-based controllers in smart buildings and avoiding disruptions in the building management system.
Handling complete short-term data logging failure in smart buildings: Machine learning based forecasting pipelines with sliding-window training scheme
Dadras Javan F.;Najafi B.;Rinaldi F.
2023-01-01
Abstract
This paper implements a machine learning(ML)-based procedure for constructing the missing sensor(s) data in a net zero energy building in case of complete failure in data recording (for up to one hour). In the first scenario, missing temperature data is re-created using the sensor's ex-ante data, the HVAC system's status flag, and the ambient conditions. In the second scenario, the temperature data (until failure occurred) from two close-by spaces are also utilized as inputs. For each scenario, ML-based pipelines' performance is first assessed by considering different prediction horizons using a benchmark algorithm. Next, each pipeline's most promising features and the most suitable algorithm are identified. Using the obtained optimal pipeline, a sliding window-based training scheme is implemented, and the size of the training window is optimized. It is shown that feature selection, algorithm optimization procedures, and the sliding window-based training scheme notably improve the forecasting performance. The proposed methodology can be deployed as a tool in intervals with total data logging failure, providing data to ML-based controllers in smart buildings and avoiding disruptions in the building management system.File | Dimensione | Formato | |
---|---|---|---|
2023 Handling complete short-term data logging failure in smart buildings Machine learning based forecasting pipelines with sliding-window training scheme official.pdf
Accesso riservato
:
Publisher’s version
Dimensione
2.78 MB
Formato
Adobe PDF
|
2.78 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.